Will SpreadConstraints make effect when replicaSchedulingType is Divided?

[vie-serendipity]

Please provide an in-depth description of the question you have:
When I test SpreadConstraints, I found it will not make effect when replicaSchedulingType is Divided. I have two clusters, where one cluster has significantly more resources than the other, and all replicas are scheduled to the larger cluster. I expect to add SpreadConstraints to make sure the smaller cluster will get at least one replica. However, it seems not both for dynamic and static.

What do you think about this question?:
I'm not sure my use case is correct. If it is, I think SpreadConstraints should be supported even replicaSchedulingType is Divided.

Environment:

• Karmada version:
• Kubernetes version:
• Others:

[XiShanYongYe-Chang]

In theory, it should be in effect. Can you share how your PropagationPolicy is set?

[vie-serendipity]

The pp is as following:

  resourceSelectors:
    - apiVersion: apps/v1
      kind: Deployment
      name: whoami
      namespace: test
  placement:
    clusterAffinity:
      clusterNames:
        - member1
        - member2
    spreadConstraints:
      - maxGroups: 2
        minGroups: 2
    replicaScheduling:
      replicaDivisionPreference: Weighted
      replicaSchedulingType: Divided
      weightPreference:
        dynamicWeight: AvailableReplicas

[vie-serendipity]

BTW, if the weightPreference is static, for example, three replicas with a weight ratio of 3:1, I believe according to the current code logic, the scheduling result would be 3:0. Would the spreadConstraint then not take effect?

I only see spreadConstraints functioning in selectclusters, not in assignReplicas. Same as dynamic.

[XiShanYongYe-Chang]

According to the pp configuration you share, there is no problem. Check whether the number of copies in the other cluster is insufficient. You can view the log of the scheduler.

BTW, if the weightPreference is static, for example, three replicas with a weight ratio of 3:1, I believe according to the current code logic, the scheduling result would be 3:0. Would the spreadConstraint then not take effect?

If it's static, it should ignore the amount of resources for allocation.

[vie-serendipity]

If it's static, it should ignore the amount of resources for allocation.

I mean static weight ratio is 3:1.

Anyway, thanks pretty much for your answer on the weekend. I will read the code further and test more.

[vie-serendipity]

After read the code, I still find SpreadConstraints doesn't work when replicaSchedulingType is Divided.

Actually, SpreadConstraints only works in selecting clusters, as following:

	clusters, err := g.selectClusters(clustersScore, spec.Placement, spec)

So it can't guarantee that the final result of assignReplicas still contains clusters of selectClusters, which means the final scheduled cluster might violate SpreadConstraints.

It's only useful when replicaSchedulingType is Duplicated.

[XiShanYongYe-Chang]

Thanks for your response @vie-serendipity
Ask @whitewindmills to help take a look and make sure.
/cc @whitewindmills

[whitewindmills]

@vie-serendipity thanks for your use case. I think it's not SpreadConstraint's fault but StaticWeight's. I mentioned a similar problem in a previous issue #4805. maybe this is a good time to help us push this feature further. @RainbowMango what do you think?